fix(retrieval): keep final release in Oxford-comma version lists

[detail-app]

Detail bug report: View on Detail

Closes #221

Bug

QueryVersionExtractor.extractVersionNumbers parses Java release tokens from learner queries for version-evidence retrieval, source-record labelling, and guided-lesson docSet scoping. Its shorthand comparison-chain regex consumed exactly one connector per lookingAt() step, so an Oxford-comma enumeration ending in , and <digit> put two connectors in sequence: the comma matched, the next token was and (not a digit), and the chain returned — silently dropping the final requested release.

• extractVersionNumbers("Java SE 21, 25, and 26") returned [21, 25], dropping 26 — one of the three indexed Oracle Java releases.
• The bug was narrow to the Oxford-comma trailing conjunction; the non-Oxford form (Java 21, 22 and 25) and the all-comma form (Java 21, 22, 25) already parsed correctly.
• Impact: the dropped release lost its dedicated docVersions=[release] retrieval constraint and its requestedVersions="..." request-anchor label across all three callers (RetrievalService, ChatService, GuidedLearningService), none of which re-derive versions downstream.

Fix

Widen the comma connector alternative in SHORTHAND_VERSION_PATTERN (QueryVersionExtractor.java:17-19) to optionally span , and in a single step:

"^\\s*(?:,(?:\\s*and\\b)?|and\\b|vs(?:\\.|\\b)|versus\\b|\\+|/|&)\\s*(\\d{1,3})\\b"

The optional (?:\\s*and\\b)? group only fires after a comma, so single-connector chains (/, vs, vs., versus, +, &, bare and) are untouched. The isPlausibleJavaRelease(>= 8) guard still absorbs low-count tails (Java 21, and 2 examples ends the chain at 2), matching the existing lone-and behavior.

Testing

• Added one focused regression-guard test (extractsOxfordCommaEnumerationWithoutDroppingFinalRelease) pinning the Oxford-comma form, a longer 4-release chain, and the two adjacent forms (all-comma, non-Oxford and) that must remain unchanged.
• Existing QueryVersionExtractorTest suite (7/7), plus the three downstream consumer suites — RetrievalServiceTest (40), ChatServiceTest (7), GuidedLearningServiceCitationTest (11) — all pass with no outcome changes.
• Compile, Spotless formatting, PMD, and SpotBugs are clean.
• Ran the full CI-equivalent lane: make test (18 shell contract tests + full unit suite), make build (frontend production build + gradlew build -x test), and make lint (ast-grep Java rules, Ruby SSOT checker, frontend oxlint/eslint/svelte-check, backend SpotBugs/PMD) — all green.
• Verified end-to-end via an ephemeral JShell probe against compiled classes: extractVersionNumbers("Java SE 21, 25, and 26") now returns [21, 25, 26], and DocsSourceRegistry.javaApiDocumentationSourcesForReleases(...) now resolves 26:java/java26-complete (previously absent).

---

Automatic Fixes PRs can be configured here.